EP4300088A1 - Method for analyzing a material composition of a test object and automation component - Google Patents

Abstract

The invention relates to a method for analyzing a material composition of a test object (40), in which
- with an impedance measuring unit (5) with an electrical test signal (PS), an impedance spectrum (ImS) dependent on a frequency (f) and / or an amplitude of the test signal (PS) via a sensor (6) from the test object (40) is recorded,
- different material compositions (Mi) are stored in an assignment table (9),
- one test specimen (Pi) is provided for each material composition (Mi),
- in a detection phase (EP) for the test object (40), the respective test object (Pi) is used and with the impedance measuring unit (5) a predeterminable number (N) of test spectra (TSik) from the respective test object (Pi) are recorded and this number of test spectra (TSik) is stored as a material composition (Mi)-specific training data set (TDS) in a sample acquisition unit (10),
- in a learning phase (LP) a specific neural network (NNi) is trained for the number (N) of test spectra (TSik) of the respective material compositions (Mi),
- In an analysis phase (AP), a currently recorded impedance spectrum (ImS) of the test object (40) is analyzed with the material compositions (Mi)-specific neural networks (NNi) and the material composition (Mi) is output as a result, which is for the analyzing impedance spectrum (ImS) showed the highest agreement with the respective material composition (Mi) specific neural network (NNi).

Description

The invention relates to a method for analyzing a material composition of a test object, in which an impedance spectrum dependent on a frequency and/or an amplitude of the test signal is recorded from the test object via a sensor using an impedance measuring unit with an electrical test signal.

The invention also relates to an automation component comprising an impedance measuring unit designed with an electrical test signal to record an impedance spectrum dependent on a frequency and/or an amplitude of the test signal from a test object via a sensor.

By measuring the impedances of a test object at different frequencies, properties of the test object can be revealed. The measuring method is known as impedance spectroscopy.

There are already suitable measuring devices for impedance measurement. These allow impedance measurement at different frequencies. The resulting complex impedance values can then be represented as a curve over the frequency, namely the impedance spectrum. It is common to determine the impedance spectrum using special sensors that allow different test objects or materials to be examined.

In contrast to electrical components, where the impedances are mainly described by physical relationships and can be calculated, there are also impedance curves that cannot easily be assigned to physical laws.

It is the object of the present invention, in particular for manufacturing automation technology, to provide a method for analyzing a material composition of a test object, specifically for test objects in which the impedance curves cannot be verified by physical laws, for example by calculating a formula.

The object is achieved by a method for analyzing a material composition of a test object in that an impedance spectrum which is dependent on a frequency and/or an amplitude of the test signal is recorded from the test object via a sensor using an impedance measuring unit with an electrical test signal, wherein in a Assignment table different material compositions are stored, with one test item being provided for each material composition, the respective test item being used in a detection phase for the test object and a predeterminable number of test spectra from the respective test item being recorded with the impedance measuring unit and this number of test spectra is stored as a material composition-specific training data set in a sample acquisition unit, in a subsequent learning phase, a material composition-specific neural network is trained for the number of test spectra of the respective material compositions, in a subsequent analysis phase that is crucial for the manufacturing process currently recorded impedance spectrum of the test object is analyzed with the material composition-specific neural networks and as a result of the analysis the material composition is output, which has given the highest agreement to the respective material composition-specific neural network for the impedance spectrum to be analyzed.

In the sense of the invention, neural networks, i.e. artificial neural networks, map neuron structures with learned knowledge. From the field of artificial intelligence there is one Knowledge discovery in database process, which is often also referred to as data mining. The general aim of discovering unknown connections from mostly very large data sets derives from this data mining. The algorithms used are operated differently, i.e. to analyze the data sets without specifying explicit result expectations. In particular, materials in automated production are subject to manufacturing or batch-related fluctuations.

The invention offers the advantage of analyzing test objects in which the relationship between object properties and impedance spectrum cannot be described analytically by physical laws.

In a further development of the method, a target material composition of the test object is stored in the allocation table and known deviations in the composition from the target material composition are selected for the other material compositions, with the target material composition being used based on the material compositions determined in the analysis phase current deviation of the material composition from the target material composition is determined.

The inventor has recognized that, in contrast to known methods of impedance spectroscopy, it is now possible to provide a material analysis in combination with an evaluation device based on AI methods in conjunction with an automation device in manufacturing technology, in particular for materials that are not physical or have chemical forms described. This has the advantage that certain impedance curves can be assigned to specific properties of the test object in a training process. A classification of the properties can then take place in real time during the ongoing production process. This classification makes one Kmo and/or monitoring of the quality characteristics of the test object possible.

In conjunction with an automation device, it is then possible to control or regulate the process. For this purpose, the method according to the invention provides, in a further development for use in a manufacturing process of the test object, to be able to intervene in the manufacturing process with the knowledge of the deviation in the material composition in such a way that the test object again has the target material composition.

An even improved method step provides that the acquisition phase is carried out on a local automation component, and the material composition-specific training data sets are sent to a higher-level computing system, with the learning phase being carried out on the higher-level computing system, and the learned or calculated material composition specific neural networks are sent back from the higher-level computing system to the local automation component and are stored in the automation component for the analysis phase.

According to the method, an automation device can then have a modular structure and even be designed as a mobile structure. This mobile structure could then only be used to train the AI applications. The data sets could then be fed into a cloud service, for example, and a trained neural network would be returned as a result. The trained neural network is then stored on the modular automation device and can then be integrated into the system in the process. A download of the neural network during operation of the system is also conceivable.

A further improvement in the significance of the analysis is achieved by the fact that the impedance spectrum is considered to be more complex Course of the frequency is specified, and separate material composition-specific neural networks are trained for an amount, a phase angle, a real part and an imaginary part, which work in parallel in the analysis phase for the comparison with the currently recorded impedance spectrum in order to ensure the accuracy of the statement to increase the material composition to be determined.

When evaluating the impedance spectrum in the form of a classification by neural networks, material compositions can be calculated in these individual forms of representation, i.e. amount, phase, real part and imaginary part, even in separate networks that work in parallel. The properties of the substance to be examined can have different effects on the individual forms of representation (amount, phase, real part and imaginary part) and this enables a more precise hit rate when classifying in the neural networks.

For an automation component comprising an impedance measuring unit designed with an electrical test signal to record an impedance spectrum dependent on a frequency and / or an amplitude of the test signal from a test object via a sensor, a processor module being designed to process neural networks, and a data memory in which a neural Network is stored, the task mentioned at the beginning is solved in that an assignment table in which different material compositions are stored, a sample acquisition unit which is designed to record a predeterminable number of test spectra for the material compositions specified in the assignment table in a detection phase by means of the impedance measuring unit are, furthermore, a data manager is designed to store and manage the number of test spectra as material composition-specific training data sets in the sample acquisition unit, an interface is designed to connect to a learning means, the learning means being designed in a learning phase for the respective training data sets to provide a material composition-specific neural network, the data memory is designed so that the specific neural networks are stored in it, the processor module being designed to make it possible to use the currently recorded impedance spectrum in an analysis phase to analyze the specific neural networks and, as a result, to output the material composition which, for the impedance spectrum to be analyzed, resulted in the highest agreement with the respective material composition-specific neural network.

The automation component is improved in that a target material composition of the test object is stored in the allocation table and the further material compositions with known deviations in the composition from the target material composition are stored, with a calculation means being present, which is designed based on the in The material composition determined in the analysis phase is used to determine the deviation of the material compositions using the target material composition.

Through the learning process, the specific properties of the test object are assigned to certain impedance curves in the later learning phase. A classification of the properties can then take place in real time during a running process. This classification makes it possible to control/monitor the quality characteristics of the test object. In conjunction with an automation device, it is still possible to control or regulate the manufacturing process.

For this purpose, the automation component is designed for use in a manufacturing process of the test object with a reporting unit, which in turn is designed to communicate the deviation in the material composition to a regulation and / or control unit, which in turn is designed in the To intervene in the manufacturing process so that the test object to be manufactured again has the target material composition.

In order to outsource the training data sets or, for example, have them processed in a cloud service, the automation component is designed with a data transfer unit, which is designed to send the material composition-specific training data sets to a higher-level computing system with the data manager.

Through a modular structure of an automation device, a mobile structure could be realized, in particular with the automation component. The mobile structure could then be solely responsible for training an AI application. The trained neural network can then later be saved back to the automation component.

FIG 1

eine Automatisierungskomponente zur Analyse einer Materialzusammensetzung,

FIG 2

den Verfahrensablauf in einer symbolischen Darstellung von unterschiedlichen Phasen und

FIG 3

eine Verdeutlichung der Ermittlung von Testspektren durch zuvor besonders präparierte Prüflinge.

The drawing shows an exemplary embodiment of the invention

FIG 1

an automation component for analyzing a material composition,

FIG 2

the process flow in a symbolic representation of different phases and

FIG 3

an clarification of the determination of test spectra using previously specially prepared test specimens.

The FIG 1 shows an automation component 1, which is designed to analyze a material composition Mi of a test object 40. The test object 40 is analyzed via a sensor 6 using a test signal PS. An impedance measuring unit 5 is designed to use the electrical test signal PS to record an impedance spectrum ImS, which is dependent on a frequency and/or an amplitude of the test signal PS, from the test object 40 via the sensor 6. Between the sensor 6 and the impedance measuring unit 5, a sensor-specific sensor interface 4 is arranged.

To evaluate a neural network NN, the automation component 1 has a processor module 7. The processor module 7 is coupled to a data memory 8, in which the neural network or networks NNi are stored.

For an acquisition phase EP (see FIG 2 ) a test object Pi is used for the test object i,O and with the impedance measuring unit 5 a predeterminable number N = 100 of test spectra TSi1,...,100 is recorded from the respective test object Pi. This number N = 100 of test spectra are stored as a material composition Mi-specific training data set TDs in a sample acquisition unit 10.

This means that in a first acquisition phase EP, a first test specimen P1, which corresponds to a first material composition Mi, is provided for measuring test spectra corresponding to this material composition. Accordingly, as many material compositions Mi are stored in the assignment table 9 as there are test specimens Pi.

For a learning phase LP, the test spectra TSi1,...,100 of the respective material composition Mi are trained to form a material composition Mi-specific neural network NNi. For this purpose, the training data sets TDs can either be guided to a learning device 12 via an interface 11 or passed on to a bus system 31 via a bus interface 32 so that the training data sets TDs can be passed on to a higher-level computing system 30. In this higher-level computing system 30 there is a learning algorithm 33, in which the training data sets TDs in material composition Mi specific neural networks NNi are trained.

A target material composition SM of the test object 40 is also stored in the assignment table 9. For the other material compositions Mi, a known deviation in the composition from the target material composition SM is also stored. There is a calculation means 13 available, which is designed to use the target material composition SM to determine the deviation of the material composition SM based on the test spectra for the material composition determined in the analysis phase AP.

Furthermore, there is a reporting unit 14 in the automation component, which is designed to communicate the deviation in the material composition to a control and/or control unit, which in turn is designed to intervene in the manufacturing process so that the test object 40 to be manufactured again has the target material composition .

According to FIG 2 The acquisition phase EP, the learning phase LP and the analysis phase AP of the method for analyzing a material composition Mi of a test object are essentially shown. In the acquisition phase EP, one test object P1, P12, P3,...,Pi-2,Pi-1,Pi is provided for the test object 40. With the impedance measuring unit 5, a number N of impedance spectra ImS are recorded for each test object Pi. The recorded impedance spectra ImS are then assigned to the test objects Pi or the material composition Mi and result in the recorded test spectra TSik of the respective test object Pi.

In a learning phase LP, a material composition Mi-specific neural network NNi is trained for the number N of test spectra TSik of the respective material composition Mi. Accordingly, a first material composition M1-specific neural network NN1 is formed from a first test spectrum TS1k for a first test item P1 and also for an i-th test item Pi for an I-th material composition An i-th test spectrum TSik is summarized and an i-th material composition-specific neural network NNi is trained or generated or formed.

In the later analysis phase AP, a currently recorded impedance spectrum ImS of the test object 40 is again analyzed with the material composition MMi-specific neural networks NNi and the result is the material composition Mi, which for the impedance spectrum ImS to be analyzed has the highest correspondence with the respective material composition Mi specific neural network NNi.

With the FIG 3 The connection between the known deviations of the test specimens Pi in the material composition from the composition to the target material composition SM is shown again. A test spectrum TS4 is recorded with a fourth test specimen, which corresponds to the target spectrum SS. A third test spectrum TS3 is then recorded with a third test specimen P3, which corresponds to the material composition SM of the fourth test specimen P4 minus a deviation D1. A test spectrum TS is recorded with a fifth test specimen P5, which corresponds to a material deviation of the fourth test specimen P4 plus a deviation D4, etc.

Claims (9)

Method for analyzing a material composition of a test object (40), in which - with an impedance measuring unit (5) with an electrical test signal (PS), an impedance spectrum (ImS) dependent on a frequency (f) and / or an amplitude of the test signal (PS) via a sensor (6) from the test object (40) is recorded, - different material compositions (Mi) are stored in an assignment table (9), - one test specimen (Pi) is provided for each material composition (Mi), - in a detection phase (EP) for the test object (40), the respective test object (Pi) is used and with the impedance measuring unit (5) a predeterminable number (N) of test spectra (TSik) from the respective test object (Pi) are recorded and this number of test spectra (TSik) is stored as a material composition (Mi)-specific training data set (TDS) in a sample acquisition unit (10), - in a learning phase (LP) a specific neural network (NNi) is trained for the number (N) of test spectra (TSik) of the respective material compositions (Mi), - In an analysis phase (AP), a currently recorded impedance spectrum (ImS) of the test object (40) is analyzed with the material composition (Mi)-specific neural networks (NNi) and the material composition (Mi) is output as a result, which is for the analyzing impedance spectrum (ImS) showed the highest agreement with the respective material composition (Mi) specific neural network (NNi). Method according to claim 2, wherein a target material composition (SM) of the test object (40) is stored in the assignment table (9) and known deviations in the composition from the target material composition (SM) are selected for the further material compositions (Mi), whereby a current deviation in the material composition is determined using the target material composition (SM) based on the material composition (Mi) determined in the analysis phase (AP). Method according to claim 2, for use in a manufacturing process of the test object (40), wherein with the knowledge of the deviation in the material composition (Mi) the manufacturing process is intervened in such a way that the test object (40) to be manufactured again has the target material composition (SM). having. Method according to one of claims 1 to 3, in which the acquisition phase (EP) is carried out on a local automation component (1), and the material composition (Mi)-specific training data sets are sent to a higher-level computing system (30), the Learning phase (LP) is carried out on the higher-level computing system (30), and the learned or calculated material composition (Mi)-specific neural networks (NNi) is sent back from the higher-level computing system (30) to the local automation component (1) and for which Analysis phase (AP) is saved. Method according to one of claims 1 to 4, in which the impedance spectrum (ImS) is specified as a complex-valued course of the frequency, and separate material composition (Mi) specific neural networks (NNi) are trained for an amount, a phase angle, a real part and an imaginary part which work in parallel in the analysis phase (AP) for the comparison with the currently recorded impedance spectrum (ImS) in order to increase the accuracy of the statement about the material composition to be determined. Automation component (1) comprising - an impedance measuring unit (5) designed with an electrical test signal (PS) to form an impedance spectrum that is dependent on a frequency and/or an amplitude of the test signal (PS). (ImS) from a test object (40) via a sensor (6), - a processor module (7) designed to process neural networks (NN), - a data memory (8) in which a neural network (NN) is stored, marked by , - an assignment table (9) in which different material compositions (Mi) are stored, - a sample acquisition unit (10), which is designed in a acquisition phase (EP)) to record a predeterminable number (N) of test spectra (TSik) for the material compositions (Mi) specified in the assignment table (9) by means of the impedance measuring unit (5). , - a data manager (2) is designed to store and manage the number of test spectra (TSik) as material composition (Mi)-specific training data sets (TDS) in the sample acquisition unit (10), - an interface (11) to a learning means (12), which is designed to provide a material composition (Mi)-specific neural network (NNi) for the respective training data sets (TDS) in a learning phase (LP), - the data memory (8) in which the specific neural networks (NNi) are stored, - an embodiment of the processor module (7), which makes it possible to analyze the currently recorded impedance spectrum (ImS) with the specific neural networks (NNK _i ) in an analysis phase (AP) and to output the material composition (Mi) as a result, which is for the impedance spectrum to be analyzed (ImS) has shown the highest agreement with the respective material composition (Mi)-specific neural network (NNi). Automation component (1) according to claim 6, wherein a target material composition (SM) of the test object (40) is stored in the assignment table (9) and the further material compositions (Mi) with known deviations in the composition to the target material composition (SM) are stored, with a calculation means (13) being present, which is designed based on the material composition (Mi) determined in the analysis phase (AP) with the help of the target material composition (SM). to determine the material composition. Automation component (1) according to claim 7, designed for use in a manufacturing process of the test object (40) comprising a reporting unit (14), which is designed to communicate the deviation in the material composition to a control and/or control unit, which in turn is designed to intervene in the manufacturing process , so that the test object (40) to be manufactured again has the target material composition (SM). Automation component (1) according to one of claims 6 to 8, comprising a data transfer unit (14), which is designed to send the material composition (Mi)-specific training data sets (TDS) to a higher-level computing system (30) with the data manager (2).

Images